Why does uniform memory consolidation sometimes degrade below the no-memory baseline?

This explores why "uniform" memory consolidation — folding every past experience into one shared store with the same compression rule — can leave an agent performing worse than if it had no persistent memory at all. The sharpest evidence sits in Does agent memory degrade when continuously consolidated?, which finds that LLM-consolidated textual memory follows an inverted-U: helpful early, then actively harmful as experience piles up. One model failed 54% of problems it had previously solved correctly once its memory was consolidated. Three culprits: misgrouping (lumping unrelated situations together), applicability stripping (saving the lesson but discarding the conditions under which it was true), and overfitting to a narrow stream of recent experience. That last one is the heart of the "below baseline" puzzle — consolidation doesn't just fail to add signal, it overwrites good default behavior with confidently-wrong generalizations.

The word doing the damage is *uniform*. The corpus repeatedly shows that the same consolidation that hurts when applied indiscriminately becomes net-positive once it's made selective or structured. Can agents compress their own memory without losing critical details? folds history into *separate* schemas — episodic, working, and tool memory — and credits exactly that structure and autonomy with avoiding the degradation that "poorly designed consolidation" suffers. Can context playbooks prevent knowledge loss during iteration? makes the same point from the opposite direction: full rewrites cause "context collapse" and brevity bias, so it uses incremental generation-reflect-curate edits instead. Uniform consolidation is essentially a repeated full rewrite — and that's the failure mode.

There's a deeper diagnosis worth knowing: forgetting may be a *misallocation* problem, not an inherent cost of remembering. Can splitting adaptation into two channels reduce forgetting? shows that routing durable lessons into fast textual context while leaving slow weights mostly untouched preserves performance and trains faster — the harm comes from cramming everything into one channel. Is long-context bottleneck really about memory or compute? and Can recurrence consolidate memory without predicting tokens? reframe consolidation as a *compute* operation done carefully during offline "sleep" passes, where more passes monotonically help — the opposite of the runaway degradation you get when you consolidate greedily and uniformly inline.

Why does uniform compression specifically push *below* the no-memory baseline rather than just to it? Because the consolidated memory becomes an input the model then over-trusts. Where do memorization errors arise in chain-of-thought reasoning? finds that local memorization — leaning on immediately preceding tokens — drives up to 67% of reasoning errors, and gets worse under distributional shift. A uniformly consolidated memory is a manufactured distributional shift: it injects stripped, misgrouped context that the model then anchors on, so the very mechanism meant to help becomes a new error source. A no-memory agent at least falls back on its clean pretrained priors.

The takeaway the corpus converges on: memory only beats baseline when consolidation preserves *applicability conditions*, keeps distinct experiences in distinct structures, and edits incrementally rather than rewriting wholesale. Strip those, and you're not adding memory — you're adding noise the model can't help but believe.